JPeetz/agent-skills / skills/marketing/local-video-production-m5
local-video-production-m5
VerifiedUse this skill when producing local videos on the 16GB M5 executor.
| File | Kind | Size |
|---|---|---|
| CHANGELOG.md | markdown | 586 B |
| SKILL.md | skill md | 9.6 KB |
| evals/evals.json | config | 2.4 KB |
| references/comfyui-mps-custom-node-pitfalls.md | markdown | 3.4 KB |
| references/kokoro-tts-mature-voice.md | markdown | 2.1 KB |
| references/ltx-mlx-a2v-setup.md | markdown | 3.9 KB |
| scripts/m5_env_check.shexecutable | script | 665 B |
references/ltx-mlx-a2v-setup.md
View on GitHub# LTX-2.x MLX Audio-to-Video (talking head) — M5 setup and usage
Primary talking-video path on 16 GB Apple Silicon: one image + one audio in →
lip-synced talking video out, MLX-native (≈2× faster than PyTorch-on-MPS), no
CUDA anywhere. Node: `noreff/ComfyUI-LTX-MLX-A2V` ("🍎 LTX MLX Audio-to-Video"
under Lightricks/MLX). Model auto-downloads on first run from
`dgrauet/ltx-2.3-mlx-q8` (Q8 quant). **On 16 GB use `dgrauet/ltx-2.3-mlx-q4`
instead** — see Memory decision below.
## Memory decision (measured on M5 16 GB, late 2026)
- Q8 @ 512×768 × 97 frames + low_ram = SWAP THRASH. 5.5 GB of 7.2 GB swap used,
~97M swapouts, ~53 min elapsed with no completion, engine at 22.5% MEM.
- Q4 @ 512×512 × 49 frames + low_ram = clean. Engine at 3.1% MEM, completes.
Measured 13:04 for 2.04 s of video = **16 s per frame**.
- Projected (at 16 s/frame): 5 s ≈ 32 min, 8 s ≈ 51 min, 22 s ≈ 2.3 h.
- The two-stage recipe is fixed: 30 guided steps at half res + 3 refine full res.
Rule: default A2V jobs on 16 GB to Q4 + 512×512 + low_ram; only consider Q8 or
taller latents on a 64+ GB machine. Interrupting dev ops: `pkill -f ltx-2-mlx`
and restart headless ComfyUI to clear stuck model state — the REST /interrupt
alone does not kill the CLI subprocess.
## ffmpeg PATH (end-of-render failure)
ltx-core-mlx encodes through a system ffmpeg. If it is not visible to the
ComfyUI process, the render runs to 100% then dies with
`RuntimeError: ffmpeg not found on PATH. Install it with: brew install ffmpeg`.
The progress bar to 100% does NOT mean the file exists — always check output.
Fix: symlink `~/homebrew/opt/ffmpeg/bin/{ffmpeg,ffprobe}` into the Comfy venv
`bin/` and export PATH in the headless launch command.
## Install (verified working)
1. Clone node into ComfyUI custom_nodes; install `requirements.txt` into the
Comfy venv (imageio, av, ltx-core-mlx, ltx-pipelines-mlx).
2. The `ltx-2-mlx` subpackage install FAILS with
`ValueError: Readme path must be within the project directory: ../../README.md`
— pyproject `readme` points outside the subdirectory and pip's PEP 517
metadata hook rejects it. A symlink into the subpackage does NOT fix it
(hatchling resolves the path before finding the local file). The fix:
clone `dgrauet/ltx-2-mlx` at the pinned tag, DELETE the `readme = ...` line
from `packages/ltx-core-mlx/pyproject.toml` AND
`packages/ltx-pipelines-mlx/pyproject.toml`, then
`pip install <clone>/packages/ltx-core-mlx` and `.../ltx-pipelines-mlx` from
the local clones.
3. `pip install imageio av "mlx-lm<0.31"` — mlx-lm 0.31.x fails to import
against transformers >=5.10, so keep transformers below 5.10.
4. Restart headless ComfyUI; verify `GET /object_info/LTXMLXA2V` returns the node.
## API workflow (from the node's own example_workflows/a2v_talking_head.json)
```
LoadImage(image) -> LTXMLXA2V(prompt, width, height, frames, frame_rate, seed,
cfg_scale=3.0, low_ram=true, model_dir="dgrauet/ltx-2.3-mlx-q8") <- LoadAudio(audio)
-> CreateVideo(images=[a2v,0], fps, audio=[a2v,1]) -> SaveVideo
```
- Audio is trimmed/zero-padded to exactly `frames / frame_rate` seconds.
- `low_ram: true` uses block streaming — mandatory on 16 GB.
- Pipeline is two-stage: 30 guided steps at half resolution, then 3 refine
steps at full resolution.
## Quality levers
- **Expressive audio drives expressive faces.** A flat/quiet track gets a
near-static mouth; the same portrait with an energetic take articulates well.
If the mouth barely moves, fix the audio before anything else.
- A start image with burned-in caption text makes the model hallucinate
gibberish captions across the whole clip — never feed a text-bearing sheet or
render as the start frame.
- >257 frames: render chunks, feeding the last frame of one chunk into the next.
- Still need a clean front-facing anchor: see the character-anchor requirements
in `creative/comfyui` references/echo-mimic-v3-flash.md.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |